fix(model): backport verification card and DiffusionGemma parity fixes to r0.7.0

[cuichenx]

Complete the combined Pruth backport to r0.7.0 with the four fixes still missing from release:

• docs(model): re-verify K-EXAONE-2 manual forward at a public revision #6474 replaces the unavailable K-EXAONE-2 manual-forward revision with its public, source-QA-verified revision and records the clamp-aware reference result.
• docs(model): re-verify DeepSeek V4 Flash GB200 performance at the current recipe #6467 updates the DeepSeek V4 Flash GB200 performance card to its recorded EP32 run.
• docs(model): re-verify GLM-5 and GLM-5.2 GB200 items at public revisions #6499 refreshes six GLM-5/5.2 GB200 SFT, PEFT and post-SFT export/inference items at a public revision.
• fix(model): disable cuBLAS TF32 in the DiffusionGemma FP32 parity scripts #6486 disables cuBLAS TF32 before computation in the four DiffusionGemma FP32 parity examples.

#6425 is now on the release branch through merged #6442. This branch merges that release update and retains the Lightning fix without a duplicate change in the PR diff. Each backport retains its source author and separate signed cherry-pick commit. A small follow-up fixes the DeepSeek test fixture to import selected recipes before patching their external dependencies, so the layout check also works when selected alone.

The card measurements remain historical results at their recorded revisions, not new release-performance measurements. DeepSeek GB300 and Nemotron Super GB300 follow-ups remain outside this change. #6474 is included unchanged from its merged source. Its author reports a next-token match at token 2879 and cosine similarity 0.999997 using a clamp-aware HF reference. This is attributed historical source QA, not independently repeated release validation. The command already required a separately produced HF logits file on the release base; the reference producer remains outside the public repository, so this PR does not make that comparison self-contained. The relevant model provider and comparison code match the recorded public revision on release. The separate inference revision remains excluded from the reachability test pending #6294.

The GLM results are also attributed historical source QA, not fresh release workload validation. #6499 preserves the pending pretrain/resume and long-context exclusions: source QA reports a GLM-5 resume OOM, an unpublished 128K dataset generator, and a missing MTP-preserving import for GLM-5.2 pretrain/resume. Those unresolved items are not repaired or validated by this PR.

The release merge resolves the overlapping revision-test file while retaining the Lightning, K-EXAONE-2 and GLM checks. Documentation tests, card-validator tests, catalog consistency and pre-commit checks were rerun after the merge. The numerical measurements below were obtained before this release merge; the DiffusionGemma implementation is unchanged, but a fresh GPU run on the updated base was not performed.

Validation on the release checkout:

• 76 documentation consistency tests and 42 card-validator tests passed; catalog regeneration and pre-commit checks passed.
• The isolated DeepSeek layout test failed before the fixture correction and passed afterward; all 13 focused recipe/precision tests passed. A negative control confirms that stale EP64 text is rejected against the real EP32 recipe.
• The Lightning, K-EXAONE-2 and GLM-5/5.2 revision checks passed locally with all checked candidate commits available. CI shallow checkouts can skip these history-dependent checks.
• On H100 with the release Bridge/Megatron-Core sources, the unchanged baseline fails the toy encoder threshold; the backport passes. Text/image max absolute errors improve from 0.01068 / 0.00570 to 0.000162 / 0.000147 (limit 0.0003), and all 82 exported tensors match exactly. TP1, TP2 and EP2 loss/gradient parity also pass with complete gradient coverage; TP1 completes an 80-step overfit check and TP2/EP2 each perform two finite updates. Fresh-process TF32, BF16 and import-behavior probes pass. All four related functional tests pass, including step-3 save/resume through step 6 with logged losses, gradients and sample counts equal to the uninterrupted run.
• The shared precision helper passes strict mypy. Strict checking across all touched upstream examples/tests remains non-clean due to existing typing and dependency-stub diagnostics; no typing cleanup is bundled.

The real 26B model and GB200 numerical checks were not rerun for this backport.

NVBugs: 6626551, 6900116, 6631267, 6875251, 6932589, 6938912.

[github-actions]

Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing:

/review

Add model=claude to use a Claude reviewer (the default is model=codex). mode=light|strict selects the review depth; for example, /review model=claude mode=strict. Comment /review help for all options.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-6512-cfc52fe5c462.docs.buildwithfern.com/nemo/megatron-bridge

[cuichenx]

/review

[svcnvidia-nemo-ci]

No verified in-scope defects found. Reviewed the card and generated-documentation updates, TF32 helper and call ordering, and affected test contracts.

This was a read-only source review; tests were not executed, and historical GPU measurements were not independently reproduced. No perf tests impacted.

Suggested test cases

• test_disable_tf32_also_disables_cublas_tf32
• test_torch_flags_alone_do_not_count_as_tf32_disabled
• test_verified_performance_layout_matches_the_published_recipe[deepseek-v4-flash-GB200-megatron.bridge.perf_recipes.deepseek] — run in isolation.
• test_lightning_card_recipes_exist_at_their_recorded_revision — use a checkout containing the recorded commits.
• test_k_exaone_2_verified_items_record_reachable_revisions
• test_glm5_verified_items_pin_reachable_revisions[glm5]
• test_glm5_verified_items_pin_reachable_revisions[glm5-2]
• test_generated_outputs_and_navigation_are_current

Note: formal approval is not available for this request, so this review is posted as a comment only.

Review generated by codex